Enable automatic column projection in groupby().agg (#12124)

This PR corresponds to the Dask-cudf version of dask/dask#9442, which was found to improve the performance of many groupby-based workflows. After this PR, 

```python
import dask_cudf

path = "/criteo-dataset/day_0.parquet"
ddf = dask_cudf.read_parquet(path, split_row_groups=10)

# The following takes <2s with this PR, and fails with
# an OOM error on main (using a 32GB GPU):
ddf.groupby("C1").agg({"C2": "mean"}).compute()
```

Authors:
  - Richard (Rick) Zamora (https://github.com/rjzamora)

Approvers:
  - GALI PREM SAGAR (https://github.com/galipremsagar)

URL: #12124

python/dask_cudf/dask_cudf/groupby.py

@@ -685,8 +685,13 @@ def groupby_agg(
             "with `sort=False`, or set `shuffle=True`."
         )
 
+    # Determine required columns to enable column projection
+    required_columns = list(
+        set(gb_cols).union(aggs.keys()).intersection(ddf.columns)
+    )
+
     return aca(
-        [ddf],
+        [ddf[required_columns]],
         chunk=chunk,
         chunk_kwargs=chunk_kwargs,
         combine=combine,

python/dask_cudf/dask_cudf/tests/test_groupby.py

@@ -130,6 +130,10 @@ def test_groupby_agg(func, aggregation, pdf):
 
     assert_cudf_groupby_layers(actual)
 
+    # groupby.agg should add an explicit getitem layer
+    # to improve/enable column projection
+    assert hlg_layer(actual.dask, "getitem")
+
     dd.assert_eq(expect, actual, check_names=False, check_dtype=check_dtype)